Week of 16 Mar 2026¶

Dates are approximate.

What happened¶

Discussion 1 (20 Mar 2026)

The discussion outlines the core architectural paradigm of Machine World, an AI-native system designed to orchestrate complex digital and physical workflows using skills, meta-skills, action templates, and Model Context Protocols (MCPs). The primary speaker details how autonomous AI agents can interface with back-end APIs, banking systems, merchant accounts, and physical hardware skeletons. To optimize cost and latency, the architecture incorporates sub-agents, Poe/Claude-like code structures, actor model cost management, and local micro-edge AI models running on mobile devices. A key technical feature discussed is the Q8 core container setup, where sidecar containers dynamically attach MCP tools on demand without drifting the agent core. Practical applications are evaluated, including automated video processing using CapCut MCP, song meaning extraction, and YouTube channel analytics via YouTube MCP. Additionally, the team plans visual story frameworks for Dhamma audio content (Marga Sakaccha) targeting Vesak releases. The session concludes with action items for audio transcription by [name withheld] and scheduling a YouTube video publication for Thursday.

Discussion 2 (20 Mar 2026)

The discussion focuses on conceptualizing and automating the production pipeline for Inner Realm music and visuals using an AI-native agent system. The speaker breaks down the dual meaning of the song 'Enne Na Wadinawa', contrasting internal mental states (tanha, avijja) with external visual stories showing Dhamma. The core requirement is to develop an interactive assistant and action model where a user can provide an audio recording and conceptual description to generate visual stories. The architecture utilizes Gemini models for conversational front-ends (especially Sinhala support), backed by an action model with skills, MCPs, and specialized agents that monitor, execute, verify, generate content, and automate channel distribution. The team aims to establish a proof of concept (POC 1) after iterating on the workflow alongside team members like [name withheld] and Bimindu.

Discussion 3 (20 Mar 2026)

In this discussion, the lead speaker outlines the system specification and architectural paradigm for the Machine World AI system. The core architecture integrates foundational omni models, Model Context Protocol (MCP) servers, action templates, execution skills, and meta-skills that dynamically generate new skills. A key immediate goal is to develop an assistive meta-skill capable of engaging with Niranjan to extract requirements and construct tailored workflows. The speaker emphasizes bridging digital intelligence with physical reality, proposing physical models and hardware interfaces capable of interacting with external machinery via Bluetooth APIs and protocol tools. The team discusses integrating real-world services—such as banking, retail platforms like Daraz, food courts, and hardware tools—via MCPs so that autonomous agents can orchestrate end-to-end tasks alongside humans. The system is also intended to support mindfulness and Dhamma retreats (Marga Sakaccha) by managing daily logistical friction and life expenses, thereby fostering mental tranquility. Miyuru and Niranjan are tasked with experimenting with a pre-configured agent core prototype to implement a concrete physical-action workflow rather than remaining purely digital. Finally, native voice interaction with custom avatars and multi-language support is highlighted as the primary modality replacing legacy web and GUI paradigms.

Key points¶

Decided · The system will be tested across two or three iterations with [name withheld] and Bimindu to achieve POC 1.
Said to exist · A YouTube MCP server is available to fetch channel analytics and facilitate automated video queries.
Planned · Define the formal system definition for the AI agent core before proceeding with model training and deployment.
Planned · Hand over recorded call audio to [name withheld] for transcription and specification extraction.
Planned · Schedule the YouTube video release for Thursday evening to allow a one-day promotion cycle.
Planned · Develop an assistive meta-skill for Niranjan to communicate with the system and extract intents to generate required workflows.
Planned · Miyuru/Miyuru must investigate how skills and MCPs can be combined together to execute user workflows within the system.
Planned · Speaker 1 will provide a pre-configured agent core so the team can start training and testing custom skills directly.

Ideas and open questions¶

Idea · Use Model Context Protocols (MCPs) as standardized tools to enable AI agents to interact with external services and physical hardware.
Idea · Connect a CapCut MCP server to the system backend to automate video editing workflows.
Idea · Integrate an AI agent via Bluetooth to discover machine APIs and functions, exposing them through MCP tokens for natural language control.
Idea · Build MCP wrappers or tools for Sri Lankan e-commerce and retail platforms like Daraz if they do not natively support MCP.

All details (13 topics, 11 more points)

AIware Architecture: Skills, Meta-Skills, and MCP Integration: The speakers discuss structuring the Machine World system using modular skills, action templates, meta-skills, and Model Context Protocols (MCPs). This setup allows AI agents to orchestrate both digital service APIs and physical real-world actions dynamically.
Sub-Agent Workflows, Poe/Claude Concepts, and Token Cost Optimization: The call explores delegating tasks to sub-agents, using Poe code and Claude skill patterns to generate modular prompt structures. Strategies for managing token consumption and actor model cost controls are highlighted.
Dynamic Sidecar Containers, Q8 Core, and Mobile Edge Models: The lead speaker details a containerized Q8 architecture where MCP tools are dynamically loaded into sidecar containers when invoked. They also discuss running local micro-edge models on mobile devices to handle localized tasks off-cloud.
Dhamma Visual Content Automation and CapCut / YouTube MCP: The team plans visual narrative structures for Dhamma audio (Marga Sakaccha) aimed at public release during Vesak. Automated media creation is planned using CapCut MCP for video editing and YouTube MCP for publishing and channel analytics.
Creative and Dhamma Concepts of 'Enne Na Wadinawa': Speaker 1 explains that the song portrays both an internal reflection of the mind and an external visual narrative. In the song, love represents the nature of craving (tanha) and ignorance (avijja), which should be visually contrasted against the realization of Dhamma and Nibbana in physical forms.
AI Assistant and LLM Selection: The speaker proposes building an assistant that can translate audio recordings and concept descriptions into narrative visuals. For front-end conversational interaction, Gemini models (specifically mentioning Gemini Pro and Gemini 3.5 Omni models) are suggested due to their Sinhala capabilities.
Action Model, MCPs, and Agent Framework: The backend system should comprise an action model equipped with skills, MCPs, and specialized agents. The pipeline handles task execution, verification, content creation, and automated distribution/scheduling to channels, leaving only high-level creative direction to the user.
POC 1 and Iterative Testing: The speaker defines achieving this end-to-end automation workflow as POC 1. It requires iterating two to three rotations with team members [name withheld] and Bimindu to refine and prove practical usage.
Machine World Architecture: MCPs, Skills, and Meta-Skills: Speaker 1 explains the foundational layers of the Machine World system: base models, MCP tool connections, action templates, operational skills, and generative meta-skills. A meta-skill must be built first to allow Niranjan to easily specify intents and generate corresponding system workflows.
Dhamma Integration and Mental Free Space: The discussion connects system automation to Dhamma practice (Marga Sakaccha) and intensive 7- to 30-day meditation retreats where verbal fabrication (vaci-sankhara) ceases. Automating domestic routines and managing monthly expenses within a budget are intended to foster physical and mental seclusion (kaya-viveka, citta-viveka).
Model Context Protocol (MCP) and Service Integration: Speaker 1 clarifies that MCP provides tools and protocols for the AI agent to interact with digital and real-world services. Examples include connecting to Google Drive, banks, shopping services like Daraz, and physical hardware via Bluetooth or APIs.
Digital vs. Physical Agents and Real-World Action: The system distinguishes between purely digital models and physical models capable of operating machinery or skeleton interfaces. The team emphasizes moving beyond standard text-based chat to create agents that coordinate with humans and trigger tangible physical actions.
Voice-First Native Interaction Paradigm: The legacy paradigm of keyboards and screens is contrasted with native voice-driven interfaces. Modern omni models supporting native voice and multi-lingual capabilities (up to 100 languages) with customizable avatars and tones will act as the conversational front-end to the core agents.
Idea · Deploy the core system inside a Q8 container framework that dynamically attaches sidecar MCP containers based on requested user tokens.
Idea · The music video for 'Enne Na Wadinawa' should feature two parallel visual stories representing internal mind states and external form.
Idea · The system should accept an audio recording and description to generate corresponding narrative concepts through conversational interaction.
Idea · Gemini Pro / Gemini 3.5 Omni models should be used for the conversational front-end due to good Sinhala language support.
Idea · The backend must consist of an action model utilizing skills, MCPs, and dynamic agent spawning based on real-time data.
Idea · The agent framework must automate content verification and distribution/scheduling to target channels based on defined skills.
Decided · The Machine World system architecture consists of foundation models, Model Context Protocol (MCP) tools, action templates, skills, and meta-skills.
Decided · The core of the system is an agentic AI core that can be trained and fine-tuned for diverse workflows.
Decided · The system specification representation is not standard imperative code, but rather an insightful prompt containing structured skills.
Planned · Niranjan and Miyuru must collaborate on a concrete prototype involving a physical action rather than purely digital interactions.
Decided · The user-facing conversational interface should use native voice models supporting around 100 languages, with customizable voice tones and avatars.